Taygete at SemEval-2022 Task 4: RoBERTa based models for detecting Patronising and Condescending Language

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 496 - 502 July 14-15, 2022 ©2022 Association for Computational Linguistics Taygete at SemEval-2022 Task 4: RoBERTa based models for detecting Patronising and Condescending Language Jayant Chhillar C2FO Noida, India IIIT-Delhi Delhi, India jayant17154@iiitd.ac.in Abstract This work describes the development of dif- ferent models to detect patronising and conde- scending language within extracts of news arti- cles as part of the SemEval 2022 competition (Task-4). This work explores different models based on the pre-trained RoBERTa language model coupled with LSTM and CNN layers. The best models achieved 15th rank with an F1-score of 0.5924 for subtask-A and 12th in subtask-B with a macro-F1 score of 0.3763. 1 Introduction The use of Patronising and Condescending Lan- guage (PCL) in text or speech can affect healthy communication channels adversely. The effect of PCL on the vulnerable sections of society have been widely studied. PCL acts as a catalyst for dis- criminatory behaviour (Mendelsohn et al., 2020) against various vulnerable groups. It has been ob- served to promote exclusion and discrimination among communities and provide a conducive en- vironment for rumour spreading and misinforma- tion (Nolan and Mikami, 2013). These negative effects of PCL are unaffected by the intent of the writer/speaker who might have unknowingly used PCL. These reasons provide a strong argument for developing methods that can identify and prevent unwanted use of PCL in news articles, blogs, and other pieces of text. In subtask-A of the Patronizing and Conde- scending Language Detection task at Semeval-2022 (Pérez-Almendros et al., 2022), the goal is to de- velop a model which takes a sample text as an input and outputs a label indicating the presence or ab- sence of PCL. In subtask-B, the model was required to identify the correct set of PCL categories. The model takes in a sample text as an input and return seven separate outputs each indicating the presence or absence of the pre-defined seven categories. The dataset for the task was shared by the task organis- ers in the English language. To tackle these tasks, RoBERTa based models were developed. Different variations of the models involved the use of feed- forward layers, LSTMs, CNN and their combina- tions. For subtask-A RoBERTa with LSTM, CNN and feed-forward layers outperformed all the other variations with an F1 score of 0.5924. In subtask-B RoBERTa with feed-forward layers got the best F1 score of 0.3763 as compared to the other variations. For subtask-A, this work achieved 15th rank in the leader board and 12th rank for subtask-B details of which are discussed in section 3.2. 2 Background Identification and analysis of PCL in text is well explored in linguistics (Aggarwal and Zhai, 2012), politics (Huckin, 2002), sociolinguistics (Thapar- Björkert et al., 2016) and other fields. However, in NLP it is still heavily unexplored and starting to gain traction. In the past topics such as sentiment analysis (Feldman, 2013), offensive speech identi- fication (Safaya et al., 2020) and fake news identi- fication …

Content truncated to 500 words